tfl.configs.RegularizerConfig

Regularizer configuration for TFL canned estimators.

tfl.configs.RegularizerConfig(
    name, l1=0.0, l2=0.0
)

Used in the notebooks

Used in the tutorials
TF Lattice Premade Models Shape Constraints with Tensorflow Lattice

Regularizers can either be applied to specific features, or can be applied globally to all features or lattices.

• Calibrator regularizers:

  These regularizers are applied to PWL calibration layers.

  • 'calib_laplacian': Creates an instance of tfl.pwl_calibration_layer.LaplacianRegularizer. A calibrator laplacian regularizer penalizes the changes in the output and results in a flatter calibration function.
  • 'calib_hessian': Creates an instance of tfl.pwl_calibration_layer.HessianRegularizer. A calibrator hessian regularizer penalizes changes in the slope, resulting in a more linear calibration.
  • 'calib_wrinkle': Creates an instance of tfl.pwl_calibration_layer.WrinkleRegularizer. A calibrator wrinkle regularizer penalizes the second derivative, resulting in a smoother function with less changes in the curvature.

• Lattice regularizers:

  These regularizers are applied to lattice layers.

  • 'laplacian': Creates an instance of tfl.lattice_layer.LaplacianRegularizer. Laplacian regularizers penalize the difference between adjacent vertices in multi-cell lattice, resulting in a flatter lattice function.
  • 'torsion': Creates an instance of tfl.lattice_layer.TorsionRegularizer. Torsion regularizers penalizes how much the lattice function twists from side-to-side, a non-linear interactions in each 2 x 2 cell. Using this regularization results in a more linear lattice function.

Examples:

model_config = tfl.configs.CalibratedLatticeConfig(
    feature_configs=[
        tfl.configs.FeatureConfig(
            name='age',
            lattice_size=3,
            # Per feature regularization.
            regularizer_configs=[
                tfl.configs.RegularizerConfig(name='calib_hessian', l2=1e-4),
            ],
        ),
        tfl.configs.FeatureConfig(
            name='thal',
            # Partial monotonicity:
            # output(normal) <= output(fixed)
            # output(normal) <= output(reversible)
            monotonicity=[('normal', 'fixed'), ('normal', 'reversible')],
        ),
    ],
    # Global regularizers
    regularizer_configs=[
        # Torsion regularizer applied to the lattice to make it more linear.
        configs.RegularizerConfig(name='torsion', l2=1e-4),
        # Globally defined calibration regularizer is applied to all features.
        configs.RegularizerConfig(name='calib_hessian', l2=1e-4),
    ])
feature_analysis_input_fn = create_input_fn(num_epochs=1, ...)
train_input_fn = create_input_fn(num_epochs=100, ...)
estimator = tfl.estimators.CannedClassifier(
    feature_columns=feature_columns,
    model_config=model_config,
    feature_analysis_input_fn=feature_analysis_input_fn)
estimator.train(input_fn=train_input_fn)

Args
name	The name of the regularizer.
l1	l1 regularization amount.
l2	l2 regularization amount.

Methods

deserialize_nested_configs

View source

@classmethod
deserialize_nested_configs(
    config, custom_objects=None
)

Returns a deserialized configuration dictionary.

from_config

View source

@classmethod
from_config(
    config, custom_objects=None
)

get_config

View source

get_config()

Returns a configuration dictionary.